Workplace AI risks go beyond inaccurate answers

Warwick Business School researchers have identified workplace AI pitfalls spanning over-automation, persuasive chatbot responses, uneven capabilities, reduced creativity and algorithmic bias. Research involving consultants found that retaining control and using AI selectively produced more accurate results than deeply integrating it into every sub-task. Delegating substantial work with little scrutiny delivered polished outputs but weakened accuracy, depth and skills development.

Chatbots also exhibited ‘persuasion bombing’: when challenged, they defended their original recommendations with dense explanations rather than correcting flawed analysis. Workers should treat responses as draft hypotheses, test assumptions and check consistency. In an experiment involving 758 consultants, ChatGPT users performed 40 per cent better on a creative task but were 20 per cent less likely to be right on a strategy task, despite sounding more convincing. Managers should test individual workflow components and their interdependencies before scaling deployment.

Popular-answer algorithms can create ‘ideation bubbles’ that produce similar ideas and homogeneous marketing messages. Tools designed to surface diverse concepts, alongside deliberate friction that encourages workers to question initial responses, can support innovation. Recruitment systems trained on historical decisions can also reproduce discrimination, while biased human reviewers may reinforce those flaws. Staff training, performance feedback and incentives that reward good decisions and penalise approval of poor AI recommendations can improve human judgment, even when the underlying model does not improve.